TY  - EJOU
AU  - Li, Yan-Wei 
AU  - Li, David Chunhu 

TI  - A Physics-Informed Spatial-Temporal Graph Attention Model for Traffic Forecasting and Interpretable Congestion Propagation Analysis
T2  - Computer Modeling in Engineering \& Sciences

PY  - 2026
VL  - 148
IS  - 2
SN  - 1526-1506

AB  - As urban transportation systems grow increasingly complex, accurate and interpretable traffic congestion forecasting is critical. While existing deep learning models utilize graph neural networks (GNNs) and attention mechanisms, they often struggle with physical consistency under extreme scenarios. To address this, we propose the Physics-Informed Explainable Spatial-Temporal Graph Attention Network (PI-X-STGAT). Our framework models road segments as graph nodes, integrating traffic, weather, and cyclical temporal features. The architecture comprises a Context-Aware Graph Attention Network (GAT) enhanced with Node Adaptive Parameter Learning (NAPL) for capturing dynamic spatial dependencies, a Gated Recurrent Unit (GRU) layer for temporal evolution, and a physics-inspired regularization term (<math id="mml-ieqn-1"><msub><mrow><mrow><mi>ℒ</mi></mrow></mrow><mrow><mi>p</mi><mi>h</mi><mi>y</mi></mrow></msub></math>) acting as a surrogate conservation constraint. This soft regularizer mitigates physically illogical predictions without strictly enforcing partial differential equations. Additionally, an integrated Explainable Artificial Intelligence (XAI) module extracts attention matrices to provide supportive attention-based influence analysis cues. Extensive experiments on a real-world dataset demonstrate that PI-X-STGAT exhibits competitive overall performance, achieving the lowest Jam Mean Absolute Error (MAE) among the evaluated baselines in our experimental setting while maintaining highly stable multi-step horizon forecasting. Synthetic boundary-case analyses indicate enhanced stability against extreme Out-of-Distribution (OoD) perturbations. Furthermore, the extracted attention-guided metrics offer dynamic attention-based influence analysis capabilities, providing a promising, interpretable framework for intelligent transportation decision-support.
KW  - Intelligent transportation systems; traffic forecasting; spatial-temporal graph neural network; physics-guided deep learning; Explainable AI (XAI); out-of-distribution robustness

DO  - 10.32604/cmes.2026.086216